Generalisation and Robustness of Deep Learning Methods by Exploiting Structural Priors
Organization:
Description:
Recent breakthroughs in Artificial intelligence (AI) allow its use as an assistive technology in healthcare. However, the adoption of AI is stalling partly because of a lack of trust by clinicians and patients. HealthyAI is about developing trustworthy and robust AI.
AI methods in medical imaging and analysis often require large training data sets and accurate labels. While the availability of large datasets continues to improve, annotation remains time-consuming, expensive, and challenging to keep cohesive. The scientific challenge is to develop AI methods that are repeatable, reproducible, and safe by requiring less training data and using problem-specific model-driven knowledge inside. This project will develop repeatable, reproducible, and safe physics-informed graph neural networks by including model-driven data symmetries and manifold-based graph embeddings.
As a consequence of this study, we will investigate mechanisms to exploit structure in data and make conclusions regarding robustness for these methods. Furthermore, this project will analyse the relationship of generalization, expressivity and robustness that is directly linked to the structure (geometry) of the data.
Output:
Joint Manifold Learning and Optimal Transport for Dynamic Imaging (2025)In Scale Space and Variational Methods in Computer Vision - 10th International Conference, SSVM 2025, Proceedings (pp. 400-411) (Lecture Notes in Computer Science; Vol. 15667 LNCS). Springer. Dummer, S., Vaish, P. & Brune, C.https://doi.org/10.1007/978-3-031-92366-1_31Data-Agnostic Augmentations for Unknown Variations: Out-of-Distribution Generalisation in MRI Segmentation (2025)[Working paper › Preprint]. ArXiv.org. Vaish, P., Meister, F., Heimann, T., Brune, C. & Wolterink, J. M.https://doi.org/10.48550/arXiv.2505.10223Data-Agnostic Augmentations for Unknown Variations: Out-of-Distribution Generalisation in MRI Segmentation - Pretrained Models and Test Data (2025)[Dataset Types › Dataset]. Zenodo. Vaish, P., Meister, F., Heimann, T., Brune, C. & Wolterink, J.https://doi.org/10.5281/zenodo.15517158 Fourier-Basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (2024)In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 17763-17772). Article 10655510. IEEE. Vaish, P., Wang, S. & Strisciuglio, N.https://doi.org/10.1109/CVPR52733.2024.01682Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (2024)[Working paper › Preprint]. ArXiv.org. Vaish, P., Wang, S. & Strisciuglio, N.https://doi.org/10.48550/arXiv.2403.01944Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification: CIFAR-10 models (2024)[Dataset Types › Dataset]. Zenodo. Vaish, P., Wang, S. & Strisciuglio, N.https://doi.org/10.48550/arXiv.2403.01944 Pictures:
